{"id":"96fa1b3e-1330-410f-b373-198b09e48e2b","arxiv_id":"1908.01671","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The authors introduce a novel 88-hour dataset of 'wellbeing' sounds and show that basic MFCC features with an SVM reach 57.4% accuracy on a 5-way classification task.","lead":"This paper presents a new audio dataset of over 88 hours of sounds used in healing practices, such as gongs, singing bowls, and chanting, collected from YouTube. It reports baseline computer classification results, showing that standard audio features reach 57.4% accuracy on a five-class task, with chanting performing poorly.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ASW dataset is not publicly accessible (footnote 1 only promises a link 'upon acceptance'), so the central claim of a novel, available resource is unverifiable.","rationale":"The paper is a standard dataset-and-baseline contribution. The strongest evidence is the detailed curation pipeline and the use of open toolkits (librosa, scikit-learn, youtube-dl), but the dataset itself is not yet accessible. The reader's verdict of CONDITIONAL is appropriate; my stress-test identifies the unavailability as the single most load-bearing concern because it precedes every other evaluation. If the dataset were released, the label-quality concern (no inter-annotator agreement) and the chanting-column inconsistency in Figure 2(d) would be the next items to check. I do not find fatal flaws in the experimental protocol: the train/val/test split is by video, which avoids segment-level leakage, and UAR addresses class imbalance. The 5-class result of 57.4% UAR is modest but plausible given the challenging acoustic similarity among gong, singing bowl, and chimes. Therefore, the verdict should remain CONDITIONAL pending public release of the data. I partially agree with the reader's weakest_assumption: the reader chose label accuracy, whereas I consider dataset availability the more fundamental gate; the reader did note the link placeholder in the rationale.","tokens_in":8498,"tokens_out":4770,"duration_ms":47859,"concrete_test":"Check whether a public, downloadable ASW dataset exists at the URL promised in footnote 1 (or any successor repository) as of the current date; attempt to download at least one audio segment from each of the five classes and confirm its file name, class label, and duration against Table 1 and Section 2.3. If no link or files are found, the paper's availability claim and baseline numbers remain unverified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 states that 'The full ASW dataset is available to the research community' but footnote 1 reads 'link will be added upon acceptance.' In the posted v2, no URL or repository is provided. The central claim of this paper is the dataset itself; without a public download, the claimed 68 clips, 5,283 60-second segments, 88:03 total duration, class distributions, and the MFCC 5-class UAR of 57.4% cannot be independently confirmed or used. This is load-bearing because the dataset is the deliverable, not an auxiliary artifact. A secondary issue is label quality: the manual listening step in Section 2.2 has no inter-annotator agreement, and the chanting class was added from the drumming search results, so label correctness is uncertain. However, even if labels are sound, the unavailability blocks the contribution. Additionally, Section 4 states the chanting class is 'completely mis-classified,' but Figure 2(d) shows 18 correct predictions for that class, indicating a possible misreading of the confusion matrix; this further argues for releasing data and code for audit.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Acoustic Sounds for Wellbeing (ASW) dataset, 88+ hours of YouTube-derived audio in five classes (Gongs, Drumming, Chimes, Singing Bowls, Chanting), together with baseline classification experiments using MFCC and Deep Spectrum features with a linear SVM. The authors report unweighted average recall (UAR) up to 57.4% for the full five-class task, noting that the chanting class is difficult. The central claims are that the dataset is a novel resource for sound-healing research and that the baselines provide a reference for future work.","tokens_in":8686,"tokens_out":5984,"duration_ms":58084,"significance":"If the dataset is made publicly available and the reported numbers are verified, the ASW dataset could be a useful community resource for soundscape analysis, audio generation, and wellbeing-related audio research. The baseline classification results, while modest, provide a reasonable starting point for future methods. The paper's use of a held-out test partition is appropriate. However, the contribution is currently not verifiable because the dataset itself is not accessible, and several internal inconsistencies in the data description prevent the reader from trusting the reported statistics.","major_comments":[{"comment":"The dataset is the central deliverable of this paper, but it is not available: footnote 1 says 'link will be added upon acceptance' and the posted version provides no URL or repository. Without public access, the claimed 68 clips, 5,283 segments, 88:03 total duration, and the classification results cannot be independently confirmed or used by the community. A working download link and a detailed description of the release (file formats, directory structure, metadata, licenses) must be provided before the dataset claim can be accepted.","section":"Section 2, footnote 1"},{"comment":"Table 1 contains an internal inconsistency that undermines the data description. The class rows sum to 34+35+35+34+15 = 153, while the total number of videos is stated as 68 in the '# videos' row; the column sums of the class rows (55, 49, 49) also do not match the partition sizes (24, 22, 22). Either the row/column meanings are mislabeled, or the entries are for a different quantity (e.g., number of segments), but as printed the table is self-contradictory. This needs to be corrected and clarified.","section":"Table 1"},{"comment":"The text states that 'for the 5-class classification (experiment 4), the chanting (cha) class is completely mis-classified,' but the confusion matrix in Figure 2(d) shows 18 correct predictions for the cha row, corresponding to a per-class recall of 18/68 ≈ 26.5%, which is above the 20% chance level. The claim of 'completely mis-classified' is therefore factually incorrect and should be revised to a more accurate characterization, such as 'poorly classified with recall near chance-level but not zero.'","section":"Section 4, Figure 2(d)"},{"comment":"The label quality is load-bearing for the dataset's validity, yet the manual listening step is described only as 'listened over briefly' to check 'quality and presence of any additional audio.' No inter-annotator agreement, explicit inclusion/exclusion criteria, or thresholds (e.g., what counts as 'excessive speech') are provided. Since the chanting class was created from drumming search results and the paper reports that speech could not be fully removed, the authors should provide a clearer annotation protocol or at least a quantitative estimate of label purity to support the dataset's use as ground truth.","section":"Section 2.2"}],"minor_comments":[{"comment":"The abstract says '5-classes of acoustic instrumentation (Gongs, Drumming, Singing Bowls, and Chanting)' but lists only four classes; Chimes is missing and should be added.","section":"Abstract"},{"comment":"The 'Dev.' column in Table 2 is not clearly defined; the text says complexity is selected on the validation set, but it is not stated whether Dev. reports validation performance of the selected model before retraining on the concatenated train+validation sets. The large gaps between Dev. and Test values (e.g., 34.8% vs. 80.5% in experiment 2) warrant an explicit explanation.","section":"Section 3.3 and Table 2"},{"comment":"The Deep Spectrum extraction is insufficiently specified for reproduction: the paper should state the exact version of the toolkit, the input representation details (e.g., spectrogram window size and hop length), and whether the VGG19 network was used with ImageNet weights or with any fine-tuning.","section":"Section 3.2"},{"comment":"The confusion matrices are not annotated with axis labels; the authors should indicate which axis corresponds to true class and which to predicted class.","section":"Figure 2"},{"comment":"The crawling date, the total number of videos downloaded, and the number of videos discarded at each post-processing step are not reported. Providing these numbers would help assess the data collection process.","section":"Section 2.1"},{"comment":"There are minor language and typographical errors, including 'significant affect' in the introduction (should be 'effect'), 'affective' in the conclusion (should be 'effective'), and 'reptitious' for 'repetitive' in the description of drumming. A careful proofread is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"This is a short dataset paper whose main contribution is the resource itself. The technical approach is standard and the baseline experiments are plausible, but the unavailability of the dataset, the inconsistency in Table 1, and the misreading of the confusion matrix are serious issues that must be fixed. If the authors can provide a working link and correct these errors, the paper would be suitable for publication; otherwise, the contribution cannot be verified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Name],\n\nThe one thing to know about this paper is that its contribution is the ASW dataset, and the dataset is not actually available anywhere. The text says the full dataset is available to the research community, but footnote 1 says the link will be added upon acceptance. So the central claim is unverifiable as posted.\n\nThat said, the paper has real merits. The dataset seems to be the first public collection aimed at acoustic wellbeing sounds, with 88+ hours across five classes (gongs, drumming, chimes, singing bowls, chanting) gathered from YouTube under Creative Commons. The collection and post-processing are described in reasonable detail: they listened through files to remove unsuited audio, chunked into 60-second segments, normalized, and removed silence. The class descriptions cite relevant wellbeing literature, which is useful context. The baseline experiments are standard but sensible: SVM with MFCC and Deep Spectrum features, four class configurations, UAR as the metric because of imbalance. The finding that conventional MFCC beats Deep Spectrum on this task is worth reporting.\n\nThe soft spots are mostly around claims outstripping the evidence. The missing link is the big one. Second, label quality is uncertain: the manual selection used only one listener, no inter-annotator agreement, and the chanting class was added from the drumming search results, so its purity is unclear. Third, Section 4 says the chanting class is 'completely mis-classified,' but the confusion matrix in Figure 2(d) shows 18 correct chanting predictions out of 68 test segments (26.5% recall). That is low, but not 'completely mis-classified.' The overstatement should be corrected. Minor: no error bars or repeated runs, but that is acceptable for a dataset baseline.\n\nWho is this for? Researchers working on soundscape design, audio generation, or wellbeing audio. If the dataset is released, this becomes a useful resource paper. As posted, it is a description of a dataset that no one can use yet.\n\nMy recommendation: send it to peer review, but make the decision conditional on the dataset actually being available and on fixing the chanting overstatement. The underlying idea is sound, and the paper is honestly written apart from that one confusion-matrix phrasing.","headline":"The ASW dataset is a plausible first for acoustic wellbeing audio, but the paper's central claim is unverifiable because the data is not actually available, and the 'completely mis-classified' chanting statement conflicts with the paper's own confusion matrix.","tokens_in":9190,"tokens_out":2973,"would_cite":false,"duration_ms":27904,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces a new 88-hour dataset of healing sounds and a 57.4% MFCC baseline for classifying them.","keywords":["dataset","sound healing","acoustic instruments","classification","MFCC","support vector machine","wellbeing","YouTube audio"],"falsifier":"Take a random sample of chunks and have three independent annotators label each by ear; if agreement falls far below the level implied by the baseline, or if a large share of chunks is reassigned, the labels and the 57.4% result would not stand.","tokens_in":8297,"feed_emoji":"🎵","tokens_out":4212,"duration_ms":40049,"temperature":0.7,"pith_summary":"The paper establishes the Acoustic Sounds for Wellbeing (ASW) dataset as a new resource for studying audio that is thought to promote wellbeing. It gathers 88+ hours of YouTube audio into five classes: Gongs, Drumming, Chimes, Singing Bowls, and Chanting. The authors also run a first baseline classification and report that conventional MFCC features with a linear support vector machine reach 57.4% unweighted average recall on the full five-class task. This matters because no comparable dataset was publicly available, and a standard benchmark gives future work on soundscape augmentation and audio generation a point of comparison.","feed_headline":"Healing-sound dataset yields a 57.4% baseline","feed_subtitle":"Eighty-eight hours of gongs, drums, chimes, singing bowls, and chanting give audio researchers a first benchmark.","key_machinery":"The carrying object is the ASW dataset itself: 68 video-derived audio files totaling 88:03 hours, split into 5,283 60-second chunks over three independent partitions. The classification baseline uses 20 MFCCs extracted with a standard audio toolkit and a linear-kernel support vector machine, with complexity optimized on a validation split and final results reported on held-out test chunks. A second feature set, 4,096-dimensional deep representations from a pre-trained VGG19 network, is included for comparison and performs worse, which the paper interprets as evidence that the long 60-second chunks suit conventional features better at this stage.","core_discovery":"The central claim is that the ASW dataset, collected from creative-commons YouTube videos and cleaned by brief manual listening, is a viable large-scale resource for acoustic wellbeing research. On that dataset, a linear SVM trained on 20 Mel-frequency cepstral coefficients achieves 57.4% unweighted average recall across all five classes, about 37 percentage points above chance; the same pipeline reaches 75.5% on the two-class gong-versus-singing-bowl task. The chanting class is almost entirely misclassified in the five-class setting, which the authors attribute to its small size and acoustic variability.","pith_inferences":["A testable extension would be to re-annotate a random sample of the 5,283 chunks with several listeners and measure inter-annotator agreement; high agreement would strengthen the dataset's label validity, while low agreement would call the 57.4% baseline into question.","The YouTube search strategy likely tracks a 'sound healing' genre as much as acoustic classes; separating instrument acoustics from recording style could be a follow-up experiment.","Because the paper notes that speech was difficult to remove completely, a future version could quantify how much vocal content remains and test whether classification is driven by voice detection rather than instrument identity."],"forward_implications":["Future work on classifying wellbeing audio can compare against a fixed ASW train/validation/test split instead of building ad hoc collections.","The 57.4% unweighted average recall becomes the headline benchmark; later feature sets or models need only beat it on the same partition to show progress.","The dataset is large enough to support conditional audio generation or soundscape augmentation experiments, as the authors propose for future work.","Because chanting hurts overall performance, a dedicated collection effort for that class would likely improve multi-class results."],"supporting_citations":[{"why":"Supplies the feature extraction library used to compute the 20 MFCCs for the baseline.","marker":"[37]"},{"why":"Supplies the deep image-based feature extraction pipeline used as the comparison baseline.","marker":"[38]"},{"why":"Supplies the linear support vector machine implementation used for classification.","marker":"[39]"},{"why":"Motivates MFCC features as an established choice for audio classification tasks such as speech and music.","marker":"[35]"},{"why":"Motivates MFCC features by their wide frequency coverage and fit to the natural human hearing range.","marker":"[36]"},{"why":"Supports the selection of oscillatory, repetitive instruments as healing sounds, which defines the class set.","marker":"[13]"}],"fun_headline_variants":["New healing-sound dataset: 88 hours, 5 classes, 57.4% baseline","ASW dataset offers 88 hrs of gongs, drums, bowls for sound research","Healing sounds dataset sets first benchmark at 57.4% recall","Singing bowls, chanting, drums: new dataset for acoustic wellbeing","5-class sound healing dataset yields 57.4% baseline accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The dataset's usefulness depends on the assumption that the YouTube search terms and the authors' brief listening pass correctly labelled every chunk into one of the five classes, with no independent verification.","fun_headline_variants_meta":{"raw":{"variants":["New healing-sound dataset: 88 hours, 5 classes, 57.4% baseline","ASW dataset offers 88 hrs of gongs, drums, bowls for sound research","Healing sounds dataset sets first benchmark at 57.4% recall","Singing bowls, chanting, drums: new dataset for acoustic wellbeing","5-class sound healing dataset yields 57.4% baseline accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1177,"prompt_tokens":850,"completion_tokens":327,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":224}},"tokens_in":466,"tokens_out":327,"duration_ms":3456,"temperature":1.0,"reasoning_tokens":224,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:05:45.305764+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of chunks and have three independent annotators label each by ear; if agreement falls far below the level implied by the baseline, or if a large share of chunks is reassigned, the labels and the 57.4% result would not stand.","supporting_citations":[{"cited_title":"Altered states during shamanic drumming: a phenomenological study,","cited_arxiv_id":null,"evidence_quote":"Supplies the feature extraction library used to compute the 20 MFCCs for the baseline."},{"cited_title":"The drum is the shaman, the spear guides his voice,","cited_arxiv_id":null,"evidence_quote":"Supplies the deep image-based feature extraction pipeline used as the comparison baseline."},{"cited_title":"The role of altered states of consciousness in native american healing,","cited_arxiv_id":null,"evidence_quote":"Supplies the linear support vector machine implementation used for classification."},{"cited_title":"The Paiste Sound Creation Gongs in Music Therapy,","cited_arxiv_id":null,"evidence_quote":"Motivates MFCC features as an established choice for audio classification tasks such as speech and music."},{"cited_title":"Life rhythm as a symphony of oscillatory patterns: electromagnetic energy and sound vibration modulates gene expression for biological signaling and healing,","cited_arxiv_id":null,"evidence_quote":"Motivates MFCC features by their wide frequency coverage and fit to the natural human hearing range."},{"cited_title":"The role of noise in clinical environments with particular reference to mental health care: A narrative review,","cited_arxiv_id":null,"evidence_quote":"Supports the selection of oscillatory, repetitive instruments as healing sounds, which defines the class set."}],"review_version":1}